fireviewer/rf-detr-large-ground-fire-smoke-v2

FireViewer RF-DETR Large Ground Fire/Smoke v2

RF-DETR Large trained to detect visible flames and visible smoke in ground-view imagery. The release contains an ONNX model ready for inference and the selected full PyTorch checkpoint. It is not an adapter and does not require a merge step.

Classes

1. flame_visible
2. smoke_visible

Validation

• EMA mAP@50: 0.70478421
• EMA mAP@50:95: 0.43243954
• mAP@50: 0.69583553
• mAP@50:95: 0.42541423
• F1: 0.65890604
• Precision: 0.71404356
• Recall: 0.61310589

The run completed 3 epochs and 4142 optimizer steps. Detailed metrics are available in metrics.json.

ONNX inference

pip install -r requirements.txt
python inference_onnx.py --model rfdetr-large.onnx --image image.jpg --threshold 0.30

The ONNX graph uses opset 17, a dynamic batch dimension, and a fixed spatial input of 512x512. It returns normalized boxes and class logits; the companion script performs the matching preprocessing and postprocessing.

PyTorch loading

from huggingface_hub import hf_hub_download
from rfdetr import RFDETR

checkpoint = hf_hub_download(
    repo_id="fireviewer/rf-detr-large-ground-fire-smoke-v2",
    filename="checkpoint_best_total.pth",
)
model = RFDETR.from_checkpoint(checkpoint, device="cuda")

Use device="cpu" on a machine without CUDA. Only load PyTorch checkpoints from trusted sources.